nklockiewicz/react-native-executorch-demo-models

react-native-executorch demo models

ExecuTorch exports of models used in react-native-executorch demo apps.

moebius-eraser

Moebius (Duan and Xu et al., ECCV 2026), a 0.22B latent-diffusion inpainting model from HUST and VIVO AI Lab, exported to ExecuTorch for on-device object removal.

file	backend	precision	size
moebius-eraser/coreml/moebius_vae_encoder.pte	CoreML	fp16	69 MB
moebius-eraser/coreml/moebius_unet.pte	CoreML	fp16	447 MB
moebius-eraser/coreml/moebius_vae_decoder.pte	CoreML	fp16	100 MB
moebius-eraser/xnnpack/moebius_vae_encoder.pte	XNNPACK	fp32	137 MB
moebius-eraser/xnnpack/moebius_unet.pte	XNNPACK	fp32	905 MB
moebius-eraser/xnnpack/moebius_vae_decoder.pte	XNNPACK	fp32	198 MB

The three graphs are meant to be driven by a DDIM sampling loop on the JavaScript side.

• moebius_vae_encoder: (1,3,512,512) image in [-1,1] -> (1,4,64,64) latents, already scaled by 0.13025
• moebius_unet: (noisy_latents, mask64, masked_latents, timestep, guidance) -> (1,4,64,64) predicted noise. Classifier-free guidance is baked into the graph as an internal batch of 2, so one call per denoising step.
• moebius_vae_decoder: (1,4,64,64) latents -> (1,3,512,512) image in [0,1]

Spatial dimensions are static at 512x512 (64x64 latent): the cross-attention positional embedding is a learned parameter tied to the training resolution.

Reference implementation, export script and sampler: on-device-eraser demo app.

License and credit

Moebius weights and code are Apache 2.0, and these exports inherit that license. The VAE comes from PixelHacker by the same authors.

@inproceedings{DuanAndXu2026Moebius,
      title={Moebius: 0.2B Lightweight Image Inpainting Framework with 10B-Level Performance},
      author={Kangsheng Duan and Ziyang Xu and Wenyu Liu and Xiaohu Ruan and Xiaoxin Chen and Xinggang Wang},
      year={2026},
      booktitle={European Conference on Computer Vision (ECCV)}
}

da3-scanner

ExecuTorch exports of Depth Anything 3 (ByteDance Seed, ICLR 2026) any-view models, for the on-device 3D scanner demo app: 8 photos in, per-view depth, confidence and camera poses out in one forward pass, no SfM and no per-scene optimization.

Layout: da3-scanner/coreml/ (iOS) and da3-scanner/xnnpack/ (Android and CPU). All graphs are static at 8 views of 560x420 (portrait), ImageNet-normalized RGB, input (1,8,3,560,420):

• da3_base_8v_560x420_coreml_fp16_gpu.pte (480MB): DA3-BASE, fp16, compute units pinned to CPU_AND_GPU. The pin is deliberate: the ANE compiler crashes the host process on this graph, while the GPU runs it in ~3.1s on an iPhone 16 Pro.
• da3_small_8v_560x420_coreml_fp32.pte (382MB): DA3-SMALL, ~2.0s on the same phone, ANE-safe.
• da3_base_8v_560x420_xnnpack.pte (413MB) and da3_small_8v_560x420_xnnpack.pte (105MB): fp32 CPU builds.

Outputs per view: depth (1,8,560,420), confidence (1,8,560,420), and a 9-float pose encoding (translation, quaternion xyzw, vertical/horizontal FOV) relative to view 0. Export script and the pose decoding: see the demo repo.

DA3-SMALL and DA3-BASE weights and code are Apache 2.0, and these exports inherit that license.

@article{depthanything3,
      title={Depth Anything 3: Recovering the visual space from any views},
      author={Haotong Lin and Sili Chen and Jun Hao Liew and Donny Y. Chen and Zhenyu Li and Guang Shi and Jiashi Feng and Bingyi Kang},
      journal={arXiv preprint arXiv:2511.10647},
      year={2025}
}

sam3

SAM 3 (Meta, 848M parameters), text-prompted instance segmentation ("segment every X"), exported to ExecuTorch for the on-device SAM 3 demo app. Type a word, every instance is segmented on the phone.

Layout: sam3/coreml/ (iOS). All graphs are static at a 672x672 input (SAM 3 natively uses 1008; 672 finds the same instances on 6 of 7 test prompts and keeps the ViT on the Neural Engine). Core ML fp16.

file	runs	size	iPhone 16 Pro
sam3/coreml/sam3_vision0_672_coreml.pte	per photo	230 MB	~370 ms
sam3/coreml/sam3_vision1_672_coreml.pte	per photo	224 MB	~330 ms
sam3/coreml/sam3_vision2_672_coreml.pte	per photo	224 MB	~330 to 500 ms
sam3/coreml/sam3_vision3_672_coreml.pte	per photo	238 MB	~330 ms
sam3/coreml/sam3_head_672_coreml.pte	per prompt	105 MB	~170 to 300 ms
sam3/coreml/sam3_text_672_coreml.pte	per prompt	707 MB	~11 to 20 ms
sam3/coreml/sam3_tokenizer.json		4 MB

• sam3_vision0..3: the 32-layer ViT in four 8-layer chunks (the ANE compiler rejects the whole ViT as one program). vision0 takes (1,3,672,672) pixels in [-1,1] (x/127.5 - 1, photo stretched to square) and returns (1,48,48,1024); vision1 and vision2 map that to the same shape; vision3 adds the FPN neck and returns 3 maps (1,256,192,192), (1,256,96,96), (1,256,48,48).
• sam3_text: CLIP text encoder, input_ids and attention_mask (1,32) int64 (CLIP tokenizer, padded with 49407) -> (1,32,256).
• sam3_head: (fpn0, fpn1, fpn2, text_features, attention_mask) -> masks (1,200,192,192) logits, boxes (1,200,4) normalized cxcywh, logits (1,200), presence (1,1). Score = sigmoid(logit) * sigmoid(presence). FPN position encodings are baked in.

Verified end to end against PyTorch on real photos: identical instances and scores, mask IoU 0.994 to 1.0.

These exports are derivatives of the SAM 3 weights and are distributed under Meta's SAM License, a copy of which is in sam3/LICENSE. They are not covered by the Apache 2.0 license of the other folders.

@misc{carion2025sam3segmentconcepts,
      title={SAM 3: Segment Anything with Concepts},
      author={Nicolas Carion and others},
      year={2025},
      eprint={2511.16719},
      archivePrefix={arXiv}
}